LlamaFactory fine-tunes 100+ LLMs, first LLM company IPO, and SHAP for model explainability.
xAI is building a 2GW data center in Mississippi, a $20 billion bet on future compute demand that impacts infrastructure planning. Yesterday also saw the first LLM-focused company go public, providing market context for builders and founders. If you are fine-tuning, a new tool unifies efficient training for over 100 LLMs and VLMs.
Elon Musk's xAI is investing over $20 billion to build a massive 2GW AI data center, named "MACROHARDRR," in Southaven, Mississippi. The governor called this project the largest economic development in Mississippi history, with operations set to begin in February 2026.
Read full story→Z.ai, the international brand for Beijing Zhipu Huazhang Technology (commonly called Zhipu AI) and developer of the GLM family of LLMs, became the world's first large language model company to go public, debuting on the Hong Kong Stock Exchange. The company raised a $558m IPO, achieving a market capitalization of ~US$6.6b to US$6.8b (HK$51b to HK$52.8b) on its debut, focusing on open-weight LLMs primarily in China.
xAI is prepping "Grok Build," a new "vibe coding agent" that will launch as a local tool with a CLI. Early details show it integrates via an npm package and websockets, offering deep configuration for environments, repos, and GitHub. xAI is betting on secure, local execution and developer-centric workflows.
A survey reviews controllable generation with text-to-image (T2I) diffusion models, detailing theoretical foundations and how novel conditions integrate into the denoising process. Use its taxonomy to pick your control handle: structure (pose, depth), identity (subject), style, or edit-based control. The curated GitHub repository is the quickest 'map' of what to implement.
This article reveals AI coding assistants are LLM-tool conversation loops, achievable in ~200 lines of Python. The orchestration loop is small, but production agents need sandboxing, tool permissioning, retrieval, and eval harnesses. It guides building an agent with `read_file`, `list_dir`, and `edit_file`, showing how the LLM dynamically uses tools based on prompts.
Apple ML's new AdaBoN strategy makes Best-of-N alignment for language models cheaper and faster. It uses a two-stage, prompt-adaptive algorithm to dynamically allocate compute, cutting down on inference-time expense.
Validation-only chatbots can regress outcomes. If you ship 'therapy' UX, you need crisis escalation, refusal policies, and measurable outcome loops.
A new analysis argues that the Model Context Protocol (MCP) is a security liability and architectural mess. It points to incoherent toolboxes, opaque resource management. Recent reports and CVEs highlight issues like prompt hijacking, DNS rebinding risks, and filesystem containment bypass in some MCP tooling. Treat MCP servers as high-privilege surfaces.
LlamaFactory is an open-source framework for efficiently fine-tuning over 100 large language models (LLMs) and vision-language models (VLMs). Presented at ACL 2024, it streamlines adapting pre-trained models for specific tasks and applications.
SHAP (SHapley Additive exPlanations) is a Python library that applies game theory to explain machine learning model predictions. It assigns an importance value to each feature, showing why a model made a specific decision as feature attributions (exact for supported tree ensembles, approximate for others).
PokerBench is a new community benchmark with simulator and leaderboard that pits LLMs against each other in poker, running ~20,925 aggregated hands (heads-up) to test strategic reasoning. It provides a leaderboard comparing models like Gemini 3 Flash and Opus 4.5 on win rates, profit, and associated costs.